Research On Improved SSD Algorithm For Vehicle Object Detection

Vehicle object detection is one of the key technologies in the intelligent transportation system.Accurately identifying and accurately locating all vehicle helps the subsequent tasks.Object detection based on deep learning is an important means to achieve vehicle object detection.SSD(Single Shot Multi Box Detector),as one of the best accuracy and speed trade-off detection algorithm based deep learning,has been widely used in a large number of real scenes.However,using the existing SSD directly for vehicle object detection,the performance has not yet fully met the demand.Therefore,it is of great theoretical and practical significance to study an improved SSD algorithm suitable for vehicle targets to improve its detection accuracy and thus meet the practical application requirements of vehicle detection.Based on the in-depth analysis of the detection principle of the SSD algorithm,the paper conducts SSD method research around three aspects of multi-scale feature pyramid,cascade detection and network acceleration,and builds a complete improved SSD algorithm framework named Cas-FESSD.The main work and contributions of the paper are as follows:(1)The thesis analyzes the principle of SSD algorithm in detail from the perspective of feature pyramid,prior frame mechanism,loss function and so on.Aiming at the problem of poor performance of low-level feature detection on small-scale vehicle targets,the paper proposes a feature enhancement strategy FES.Based on the introduction of high-level features with sufficient semantic information,an additional parallel shallow network is designed to supplement the missing vehicle details,and Use the extended receptive field module based on dilation convolution to effectively increase the receptive field,thereby accurately identifying small-scale vehicle targets.(2)In order to improve the positioning performance of the algorithm,the paper introduces a cascade detection strategy to the SSD algorithm.Aiming at the misalignment of the features and anchors(Anchor)detected by the cascade SSD algorithm for the second detection,the paper proposes a feature alignment module FAM based on the deformable convolution structure,which uses the regression of the previous detection to the Anchor space position to generate the volume Multiply the sampling point deviation,thus alleviating the Anchor position offset problem.(3)In response to the change in the input distribution of the second detection of the cascade SSD,the paper divides the positive and negative sample thresholds as the entry point,considers each vehicle object separately,and calculates the distribution of the cross ratio of all positive samples and the target frame on the basis of the original threshold situation,and then propose a sample independent incremental adaptive positive and negative sample threshold determination method,thereby changing the distribution of training samples.(4)Aiming at the problem that the characteristics of different positive samples have different expression capabilities on the object and affect the performance of regression performance,the paper calculates the "Centerness distance" between the center point of the Anchor and the center point of the vehicle as the weight term in the regression loss function.The regression loss is weighted to avoid features that are far from the target because it contains more background noise and thus brings the wrong gradient descent direction to the regression.(5)Considering that the additional convolution and other structures will have a certain impact on the inference speed of the SSD algorithm,this paper designs a strategy for pooling layer pruning on the backbone network,and uses the loss function to ensure the consistency of the output of the feature map before and after clipping To minimize the accuracy of the algorithm due to cropping.Combining all the above improvements,a complete SSD algorithm framework Cas-FESSD for fast vehicle object detection is constructed.The thesis uses Udaicty data set to train and test Cas-FESSD,and sets up multiple sets of comparative experiments.A large number of experimental results show that,compared with the original SSD algorithm,the improvement strategies proposed in the paper can improve the performance of the SSD algorithm to varying degrees.At the same time,the experimental results on the UA-DETRAC data set show that Cas-FESSD is one of the best vehicle object detection algorithm with the best trade-off between accuracy and speed.